Control of bioreactor with adaptive neural network system

CONTROL OF BIOREACTOR WITH ADAPTIVE NEURAL NETWORK SYSTEM

RAJIB NAYAK

DEPARTMENT OF BIOCHEMICAL ENGINEERING &

BIOTECHNOLOGY

INDIAN INSTITUTE OF TECHNOLOGY, DELHI

APRIL, 2010

© Indian Institute of Technology Delhi (IITD), New Delhi, 2010

CONTROL OF BIOREACTOR WITH ADAPTIVE NEURAL NETWORK SYSTEM

un

RAJIB NAYAK

Department of Biochemical Engineering & Biotechnology

Submitted

in fulfillment of the requirement of DOCTOR OF PHILOSOPY

to the

INDIAN INSTITUTE OF TECHNOLOGY, DELHI

April, 2010

Dedicated to....

My Baba (Mr. Anil Kumar Nayak) and

Maa (Mrs. Sandhya Nayak)

My idol, inspiration and who gave me

every opportunity to realize and fulfill my

dreams.

CERTIFICATE

This is to certify that the thesis entitled

degree of "Doctor of Philosophy" is a record of the bonafide research carried out by him, which has been prepared under my supervision and guidance in conformity with rules and regulations of the "Indian Institute of Technology, Delhi". The results described in it have not been submitted in part or full to any other University or Institute for the award of any Degree / Diploma.

Dr. James Gomes

Associate Professor,

School of Biological Science,

Indian Institute of Technology, Delhi, New Delhi —

India.

ABSTRACT

Bioprocess engineering has seen distinct improvement in techniques of monitoring, modeling and control over the last two decades. Further, with the rapid improvement in biotechnological research has lead to the continuing development and discovery of new clinically and industrially important biomolecules everyday, there are a host of variables that always need to be determined indirectly due to lack of reliable and robust on—line hardware and / or bio—sensors. Since bioprocesses are characterized by inherently nonlinear, time varying, considerably slow in response and endure phenomenological model uncertainty, employing neural network based

"intelligent" software sensors systems is an attractive technique for innovative bioprocess control development.

The main focus of this research work concerns the search and study of ensemble of neural networks methodologies aiming at a more precise and robust estimation of state variables to develop innovative bioprocess monitoring and control strategies. In this respect, Sequential Adaptive Networks (SAN) based `soft sensor' for state estimation and one sampling step ahead prediction of key state variables is developed. SAN is an assembly of chronologically ordered single neural networks, with one sub—network assigned to each sampling interval, so that feature memory is distributed. Each of these networks in turn is adapted to changing process conditions based on on—line measurement that guide the entire ensemble along the evolution of the process. The information of the metabolic state of the process obtained from DO measurement is used to update the weights of the SAN, enabling an on—line adaptation to changing process dynamics.

Two bioprocesses, yeast production by Saccharomyces cerevisiae and methionine production by Corynebacterium lilium were employed to evaluate the performance of the SAN based control strategies. Somileitner & Kappeli model was used to describe the yeast production. The model developed for methionine production contains an exponential kinetic structure for describing the nonlinearities and metabolic switching function for describing oxygen dependency. In order to appraise the developed model, six different statistical analyses were carried out.

Three SAN based control strategies were designed for yeast production system. The control objectives were biomass concentration tracking at a predefined slope and maximization of biomass production while keeping yeast concentration at a minimum value. For methionine production process, two control objectives were defined. First control objective was methionine concentration tracking along a linear profile and a SAN assisted mechanistic feed forward controller was designed to achieve the objective. In the second control strategy, an objective function was defined to optimize the methionine production. A statistical framework was developed by combining the objective function, model equations and operational constrains on state variables to get the range of controller parameters for optimum methionine production.

A cubic data domain with different combination of state variables and parameters for each control strategy was constructed for training, validation and implementation of SAN. The SAN was trained for an average threshold Root Mean Square Error (RMSE) per data point of 10"5 for all the control strategies. The recall profiles matched the training dataset profiles with recall RMSE of the order of 10"5.

Cross validation was done to evaluate the accuracy of the SAN prediction. Data sets for validation were farthest distance from the training data sets. Results showed that

the overall Normalized Root Mean Square Error (NRMSE) for each data sets of each control strategy was in the order of 10"3 to 10"4.

In case of implementation, SAN based control strategies were tested in three different simulated process environments namely, ideal case, up to 10% noise in measured variable (DO) and process parameters uncertainty. In ideal case, results showed that overall NRMSE in the order of 10"3 to 10"4 for all the data sets with periodic adaptation of SAN. To simulate actual process conditions, the DO output of the process model was combined with 5% and 10% white noise and presented to SAN. The results showed that the prediction error was increased for both the cases and NRMSE reached in the order of 10"3 for 5% noise and 10"3 to 10 2 for 10% noise.

The parameters defining specific growth rate were randomly varied up to ±40%, to generate a truly parameter uncertain environment. The results showed that the overall NRMSE for prediction was of the order of 10"3 and the performance of SAN controlling under condition of model—plant mismatch was acceptable.

In simulated environment, the final concentration of methionine achieved for SAN feed forward controller and SAN hybrid controller was 4.8 g L"1 and 6.9 g L"1 respectively compared to 1.8 g L"1 when no control action was taken. The SAN assisted all the developed controller is steady, robust and exhibited stable tracking of control objective in different simulation environment. It is perceived that it can be used for on—line real time implementation in fed-batch bioprocesses.

ACX

r

More than six years ago, I had started my journey as a PFid) student in one of the admired institute of the country `Indian Institute of Technology, dDeOiii". The journey was not smooth ad the time but I find severa(peop(e who Helped me, guidecfine, motivatecime ancfencouraged me to finish this journey. I fee(that it is next to impossi6Ce to overcome the most valua6Ce journey of my

ft

The person, I think without whom this journey would not have been possi6Ce, my supervisor and philosopher, B~r. James Gomes. I always found7iis presence, guidance, encouragement, support and motivation when my journey was jaggec( It has been proudCprivi(ege arufunforgetta6Ce experience to work under his esteemed supervision. Sir, I am realTy grateful to you for being there for me ad the time to successfulTy complete my journey.

I wou&cfaL o Cke to thiankProf. (B~r)A. 1 Srivatava, B~r. vikram Sahcai aruf(Dr zjesh Fianna for their va(ua6Ce ac vice, guidance and support during the entire tenure of my PhD.

I want to thank alt of my (Departmental Teachers for their advise, support and affection to complete my journey. I want to also thank, other technica(staff of Process Gab ancfcDepartmental Gab. Specia(thianks forMr. D.

v

I would (ke to thank my Cab mate and my batch mate for their unconditional help and support.

SpeciaCtFianks toACok Sanjay, Ama(zndu, Anand gju, ParuC BFiawna, Anjal andAsis/i.

I am realty proucifor my friends and want to thank to Help me in the painfu(ancfdarkdays. Asif, vino, u(an, )4nup Da, Su6Fiasisf, Sa/iajahcan, 7apu, Sankar, Parthco and ?gju special thanks to a(Cyou for unconditional love and fiienc(ship.

I want to t/ianka(Cof my family members specialty Bhutum, Bu(o, Chotka and "Choto Pisemosia"

for their endless Cove and affection. I always fincfalT of you standing by my side when I neecfyolL

I ddo not know how I express my gratitude to my "cDidi" andAma(cDa. Tour Cove and affection Help me to finish my journey. Thanks to both of you for your unconditional support. "Boom", I thinkyou are my driving force. I Cove you.

Special thanks goes to my father in Caw qtr. Sisir Ghosh and mrs. ('ushpa Ghosh whoalways encouraged me.

No words are adequate enough to express my gratitude to my father qtr. An(mar Nayak ancd my mother Ctrs. Sand iiya Nayak. I think your endless Cove; affection and encouragement Help me to finish the most precious journey of my Cf. I always know that if everything goes wrong both of you are there for me to Cookafter me anct to smooth my Cf. I ddo not know how I can telT you that you are the best parents in the world dank you "BABA" anti "M% "for your affection and

Cove.

cFinalTy, I wou(~ tike to thank my wife cDipa for her endless Cove and encouragement throughout this entire journey. Worc(s fai(me to express my appreciation for the many sacrifices site has made to support me in undertaking my doctoral studies. Site was always there Cke a pillar of strength, listening to my woes andgiving me encouragement during the toughest phase of my work. I would also like to thank Fier for the countless hours I have spent with her to write this thesis. I owe an immeasura6Ce cleft and deep affection to her.

Most importantly, I would 1kç to thank the a(migkty God, for it is under his grace that we live, learn andf(ourisk.

x -6 rAraya(J

CONTENTS

Title Page

No.

List of Figures i—xiv

List of Tables xv—xv

CHAPTER 1. INTRODUCTION & OBJECTIVES 1-9

1.1 Background 1

1.2 Motivation 4

1.3 Scope and objectives 6

1.4 Significance of this Research 7

1.5 Organization of Thesis 8

CHAPTER 2. NEURAL NETWORKS ASSISTED BIOPROCESS 10-21 MONITORING AND CONTROL: A REVIEW

2.1 Introduction 10

2.2 Artificial Neural Networks 13

2.3 Application of Neural Networks in Bioprocess 17 Monitoring and Control

2.3.1 Neural Networks based Model Predictive 18 Control

2.3.2 Neural Networks based Soft Sensors 19

2.3.3 Neural Networks based Hybrid Control 20 Strategies

2.4 Conclusions 21

CHAPTER 3. MATHEMATICAL MODELING OF SEQUENTIAL 22-57 ADAPTIVE NETWORKS AND BIOPROCESSES

3.1 Introduction 22 3.2 Development of Sequential Adaptive Networks 24 3.2.1 Architecture of Sequential Adaptive Networks 24 3.2.2 Training and Adaptation Algorithm of 26

Sequential Adaptive Networks

3.2.3 Applications Methodology of Sequential 32 Adaptive Networks

3.2.3.1 Selection of Bioprocesses 32 3.2.3.2 Development of Control Strategies 35 3.2.3.3 Generation of Data Sets for Applications 36

of Control Strategies

3.2.3.4 Optimization of Sequential Adaptive 37 Networks Architectures

3.2.3.5 Training of Sequential Adaptive 38 Networks

3.2.3.6 Validation of Sequential Adaptive 40 Networks

3.2.3.7 Off—Line Implementation of Sequential 41 Adaptive Networks (Theoretical

Analysis)

3.3 Mathematical Modeling of Bioprocesses 43

3.3.1 Introduction 43

3.3.2 Fed—Batch Model for Yeast (S

.

3.3.2.1 Model Selections for Yeast Fermentation 44 3.3.2.2 Sonnleitner and Kappeli Model for Yeast 46

Fermentation

3.3.3 Fed—Batch Model for Methionine Production 54 CHAPTER 4. SAN—CONTROLLED YEAST FERMENTATION 58-108

PROCESS

4.1 Introduction 58

4.2 Sequential Adaptive Networks Assisted Feed 60 Forward Controller for Biomass Concentration

Tracking

4.2.1 Development of Feed Forward Control 61 Strategy for Biomass Concentration Tracking 4.2.2 Data Sets Generation and Implementation 62

Procedure

4.2.3 Results and Discussions 64

4.2.3.1 Training and Validation performance of 64 SAN

4.2.3.2 Performance of SAN—Feed Forward 69 Controller under Ideal Conditions

4.2.3.3 Performance of SAN—Feed Forward 73 Controller in the Presence of ±5% and

±10% Noise in DO Measurement

4.2.3.4 Performance of SAN—Feed Forward 78 Controller when Parameters are

Randomly Varied

4.2.4 Conclusions 83

4.3 Sequential Adaptive Networks Integrated Feed 84 Forward Controller for Maximization of Biomass

Production

4.3.1 Development of Feed Forward Controller for 86 Maximization of Biomass Production

4.3.2 Data Sets Generation and Implementation 88 Procedure

4.3.3 Results and Discussions 89

4.3.3.1 Comparison between Performance of 89 Heuristic Controller and Feed Forward

Controller

4.3.3.2 Training and Validation performance of 91 SAN

4.3.3.3 Performance of SAN—Feed Forward 93 Controller under Ideal Conditions

4.3.3.4 Performance of SAN—Feed Forward 96 Controller in the Presence of ±5% and

±10% Noise in DO Measurement

4.3.3.5 Performance of SAN—Feed Forward 105 Controller when Parameters are

Randomly Varied

4.3.4 Conclusions 108

CHAPTER 5. SAN—CONTROLLED METHIONINE 109-183 PRODUCTION PROCESS

5.1 Introduction 109

5.2 Sequential Adaptive Networks for Feed Forward 111 Control of L-Methionine Production

5.2.1 Development of SAN based Feed Forward 112 Control Strategy

5.2.2 Generation of Data Sets and Application 114 Procedure

5.2.3 Results and Discussions 118 5.2.3.1 Variation of Sampling Interval for 118

Determining the Number of Sub- Networks

5.2.3.2 Training and Generalization Performance 121 of SAN

5.2.3.3 Performance of SAN—Feed Forward 126 Controller under Ideal Conditions

5.2.3.4 Performance of SAN—Feed Forward 130 Controller in the Presence of Noise in

DO Measurement

5.2.3.5 Performance of SAN—Feed Forward 131 Controller with Random Variation in

Process Parameters

5.2.4 Conclusions 134

5.3 Generalized Hybrid Control Synthesis for Affine 140 System by using Sequential Adaptive Networks

5.3.1 SAN — Hybrid Control Synthesis 141 5.3.2 Design of Operating Regime 143

5.3.3 Results and Discussions 145

5.3.3.1 Generation of Data for Implementation 145 of SAN—Hybrid Controller

5.3.3.2 Training, Validation and On—Line 147 Implementation

5.3.3.3 Performance of SAN—Hybrid Controller 165 in the Absence of Measurement Noise

5.3.3.4 Performance of SAN—Hybrid Controller 171 in the Presence of ±5% and ±10% Noise in DO Data

5.3.3.5 Performance of SAN—Feed Forward 177 Controller when Parameters are

Randomly Varied

5.3.4 Conclusions 183

CHAPTER 6. SUMMARY CONCLUSIONS AND FUTURE 184-190 WORK

6.1 Summary 184

6.1.1 Mathematical Modeling of SAN and 184 Bioprocess

6.1.2 SAN—Controlled Yeast Fermentation Process 185 6.1.3 SAN—Controlled Methionine Production 186

Process

6.2 Conclusions 187

6.3 Scope for Future work 189

REFERENCES 191-207

APPENDIX A 208-237

APPENDIX B 238-249

APPENDIX C 250-260

APPENDIX D 261-266

APPENDIX E 267-269

APPENDIX F 270-279

RESUME OF THE AUTHOR